Semantics-based concurrency control: beyond commutativity

Reviewer: Margaret H. Dunham

The recoverability of operations provides more concurrency than commutativity, which in turn provides more concurrency than traditional concurrency control techniques. Conventional concurrency control techniques (such as two-phase locking) are based on the concept of conflicting operations. Two operations conflict if they both access the same data item and at least one of them is an update operation. Operations that conflict may, however, because of the specific semantics involved, be allowed to execute in parallel without harming the consistency of the database. This idea is captured by commutativity. Two operations commute if they can be executed in any order (or simultaneously) and obtain the same results. An operation y is recoverable with respect to an operation x if a user obtains the same results whether or not x executes immediately before y . Commutativity implies recoverability. Examples used throughout this paper involve operations against a set and a stack. As stated by the authors, two insert operations against a set commute and can be executed in parallel. They are, however, conflicting. Similarly, two push operations against a stack are not commutative, but they are recoverable. Both commutativity- and recoverability-based concurrency control techniques avoid cascading aborts. The major contribution of this paper is that it proposes a database concurrency control technique that is based on recoverability and thus is less restrictive than either traditional approaches or those based on commutativity. The paper contains a formal definition of recoverability and is replete with good examples illustrating it intuitively. The concurrency control strategy based on recoverability is described, its correctness is proven, and the overall algorithm for its implementation is included. The effectiveness of the algorithm is investigated by reporting on the results of simulation studies. These studies demonstrate the superiority of the proposed recoverability technique to those based on commutativity with respect to both throughput and response time. The paper is relatively easy to read, and the discussion of the simulation results is convincing. I observed several minor problems with the contents of the paper, however. One of my major concerns about using recoverability is its overhead. The paper never really addresses this issue. As stated by the authors in a personal communication, however, this additional overhead “should be balanced against number of nonblocking operations.” It would still be nice to see simulation results when overhead is considered. Another technical issue concerning the performance results is that long-lived transactions (LLTs) are not studied, although the introduction specifically states that the new technique “will be especially useful when long-lived transactions are in progress.” Again, performance results for LLTs would be nice to see. The notation used in the formal definition of recoverability is easy to follow for anyone with a good understanding of database transaction processing and particularly for those familiar with commutativity-based algorithms. The paper is recommended reading for all researchers interested in database transaction processing and concurrency control.